You are on page 1of 13

International Journal of Aeronautical and Space Sciences

https://doi.org/10.1007/s42405-020-00292-5

ORIGINAL PAPER

Intercept Point Prediction of Ballistic Missile Defense Using Neural


Network Learning
Jun‑Yong Lee1,2 · Byeong‑Un Jo1 · Gun‑Hee Moon1 · Min‑Jea Tahk1   · Jaemyung Ahn1

Received: 8 June 2018 / Revised: 6 May 2020 / Accepted: 23 May 2020


© The Korean Society for Aeronautical & Space Sciences 2020

Abstract
This study proposes an algorithm to calculate fast the predicted intercept point (PIP) of the anti-ballistic missile system. A
neural network system is trained to learn the motion of the ballistic target to predict the future target position. Then a predic-
tion algorithm iteratively calculates PIP and the launch time of the interceptor. PIP calculation enables the missile to effec-
tively approach to the ballistic target. Computer simulations using simple interceptor models are conducted, demonstrating
the usefulness of the proposed PIP determination algorithm. The proposed algorithm significantly reduces the computation
time required for target trajectory prediction in real-time.

Keywords  Predicted intercept point (PIP) · Ballistic missile defense · Neural network · Response surface method (RSM)

1 Introduction prompt PIP determination and immediate interceptor launch.


In addition, ballistic targets entering the atmosphere undergo
Ballistic missiles are strategic weapons to destroy standoff drastic velocity changes due to drag, which is very chal-
targets with high speed and high penetration. Against bal- lenging in terms of PIP determination, missile guidance,
listic missiles, various defense systems have been developed and control.
in many countries to intercept incoming ballistic missiles The ground radar starts tracking the ballistic target before
before they arrive at intended ground targets. The perfor- the launch of the interceptor. Processing the track data pro-
mance of the missile defense system is affected by the flight vides the current target states such as position and velocity.
phase of the enemy ballistic missile due to the target detect- But using this information to implement a midcourse guid-
ability of each phase [1, 2]. The boosting phase is easily ance law to steer the interceptor to the target is not recom-
detected by an infrared sensor mounted on satellites because mended because of low update rate and time delay of the
of the bright and hot exhaust plume of ballistic missiles. target information. Instead, the predicted intercept point
Detection of ballistic missiles during the coasting phase (PIP) and a suitable interceptor launch time are calculated
requires a powerful and sophisticated radar. Moreover, from the prediction of the future target trajectory and the
interception coverage is colossal in the case of the coasting interceptor’s flight time.
phase. Finally, to intercept a re-entry target, even small and Spatially, PIP is a position where the target and intercep-
light missiles are sufficient and have economic advantages. tor missile meet each other at the same time. Therefore, not
However, a re-entry target has a flight time of about 30 s, only the future position of the ballistic target but the flight
which is a lot shorter than other phases. Short flight time time of the interceptor must be known accurately. The sim-
implies that interception during the re-entry phase requires plest way to find the target future position is the integration
of the target dynamics. But this integration may take a long
calculation time if the aerodynamic effect is included for
* Min‑Jea Tahk
mjtahk@kaist.ac.kr improved accuracy. Also, the time-to-go calculation of the
interceptor using the simple rule of “range over velocity” is
1
Department of Aerospace Engineering, Korea Advanced far from being accurate because the interceptor velocity is
Institute of Science and Technology, Daejeon 34141, time-varying.
Republic of Korea
Prediction of target future position using predictor–cor-
2
Seoul National University and Automation and Systems rector has been suggested in [3] where the current target
Research Institute, Seoul, Republic of Korea

13
Vol.:(0123456789)
International Journal of Aeronautical and Space Sciences

states are estimated from the radar measurements and the and function evaluations. Using the predicted target posi-
future target position is calculated step by step using a pre- tion, PIP is determined within the effective capture region
dictor–corrector method. This method could be time con- of the defense interceptor by the proposed algorithm. The
suming while the prediction error is accumulated as the flight time data of the interceptor for the PIP candidates in
prediction interval increases. Neural network learning has the capture region is tabulated off-line in advance.
been applied for a midcourse guidance law in [4]. A simple The interceptor used in this study is composed of a sin-
iterative method is proposed to predict the future position gle-stage rocket and a kill vehicle (KV). Once the PIP is
of the target, but the motion of the target is simplified as a determined, the interceptor is launched at the launch time
rectilinear one in the simulation study. determined during the PIP calculation. After the boosting
A similar concept to PIP is the instantaneous impact point phase, the kill vehicle enters the midcourse guidance phase,
(IIP), which is the ground impact point of a sounding rocket in which the trajectory is corrected if PIP is updated. We
or a burn-out booster separated from the launch vehicle. assume that PIP is updated at every measurement of the
Accurate calculation of IIP is important for the safe opera- ground radar. Trajectory correction combined with PIP
tion of space launch vehicles and sounding rockets. There update enhances the guidance performance of the interceptor
have been several studies on IIP prediction as found in [5–7]. in response to uncertainties in target trajectory and intercept
The study of [5] has developed a linearized IIP prediction launch time. When the seeker successfully locks on the tar-
algorithm that implements first-order corrections for the get, the interceptor enters the terminal homing phase with
effect of the Earth curvature, the Coriolis force, and gravity an advanced homing guidance law.
field variation with the parabolic flight model. Keplerian In general, the ascending trajectory of the interceptor
IIP prediction algorithms using RSM (Response-Surface- is solved by trajectory optimization which is given as an
Method) are proposed in [7] where various effects on the open-loop solution. In this paper linear tangent law (LTL)
projectile motion are considered. The method of [7] does is applied in the optimization procedure during the boosting
not require an iterative algorithm unlike that of [8]. Since IIP phase. The optimization results provide the optimal LTL
is a touchdown point on the ground or sea level, its altitude parameters which are used in the practical scenario.
information is predetermined. This is the major difference This paper is organized as follows: In Sect. 2, position
between IIP and PIP. prediction of the ballistic missile using a neural network is
Calculation of the time-to-go of the interceptor also has suggested and the target trajectory is predicted to determine
an essential role in determining PIP. A simple method is the PIP. Section 3 describes in detail the guidance meth-
to approximate the time-to-go as the range over velocity. ods employed in this study for the launch phase, midcourse
More sophisticated methods are proposed in [9–11] where phase, and terminal homing phase, respectively. Engagement
the homing guidance laws are expressed as a function of the simulation of an anti-ballistic missile using the proposed PIP
time to go and a PN-based time-to-go estimate is employed algorithm against a ballistic target is conducted in Sect. 4 to
under the small-angle approximation. The study of [12] has demonstrate the practical feasibility of the proposed algo-
also suggested a recursive time-to-go estimation method for rithm. The algorithm performance is also compared with
homing guidance. This method compensates for the time-to- that of the RSM approach. Finally, Sect. 5 provides conclu-
go error in every guidance cycle for the application of PN sions and suggestions for future study.
guidance or optimal guidance. Other time-to-go estimators
for PN guidance can be found in [13–15].
This paper proposes an algorithm to determine PIP of 2 Trajectory Prediction of Ballistic Targets
the re-entry ballistic target. Contrary to other research, this
work utilizes a neural network learning to predict the future Without aerodynamic forces, a projectile is in a state of free
position of the target. The neural network system learns fall and has analytical solutions for all the state variables.
the nonlinear motion of the target and outputs a predicted In Cartesian coordinate, the position of a falling object is
location by feedforward propagation. We have also modi- expressed as a quadratic form of time. In the ECEF (Earth-
fied the response surface method (RSM) used in [7] for PIP centered, Earth-fixed), Kepler’s Equation and Hit equation
determination to evaluate the performance of the proposed [8] provide the exact trajectory. However, considering the
neural-net algorithm. As a matter of fact, PIP determination aerodynamic effect, it is hard to derive the exact solution. In
using RSM needs to designate the intercept altitude with a the troposphere of low altitude, the speed of the projectile
proper initial condition. For this reason, we choose to use decreases in proportion to the square of its speed.
the neural network method that can be applied to find the In this section, the dynamics of a re-entry ballistic mis-
PIP for arbitrary initial conditions. By virtue of the neural sile is modeled including the aerodynamic effects. Using
network structure, the future positions of the ballistic target this model, a data set of ballistic missile motion is gener-
are calculated quickly by a few additions, multiplications ated for neural network training. Then, a classical neural

13
International Journal of Aeronautical and Space Sciences

network is trained and used to predict the future trajectory 2.2 Prediction Using Classical Neural Network
of the ballistic target and the intercept point.
The equations of motion written in (1) do not permit ana-
lytical solutions. While numerical integration such as
2.1 Ballistic Target Dynamics Runge–Kutta methods can predict the trajectory precisely,
accurate integration using numerical integration requires a
In this section, the two-dimensional dynamics of a re-entry heavy computational load. This paper proposes an alterna-
ballistic target is described in the Cartesian coordinate tive approach for trajectory prediction based on neural net-
frame. The target is considered as a point mass for which work learning. ( )
the only external force is the aerodynamic force. Most The future state of the target at t = tp , 𝐱 tp , is a function
of the ballistic missiles do not have devices to maneuver of tp , the initial time t0 , and initial state 𝐱0:
during the re-entry phase. Assuming that the target has a ( ) ( )
symmetric aerodynamic configuration, we only consider 𝐱 tp = 𝐠 tp , t0 , 𝐱0 , 𝛽 . (6)
drag force given as below.
The above function{ can be } evaluated by numerical inte-
ẋ = V cos 𝛾, gration for given tp , t0 , 𝐱0  . Hence, a data set( for neural)
ẏ = V sin 𝛾, network training can be created by calculating 𝐠 tp , t0 , 𝐱0 , 𝛽
2 for wide ranges of the time span tp − t0 , initial states and bal-
𝜌gV (1)
V̇ = − − g sin 𝛾, listic coefficient. The output variables of the neural network
( )
2𝛽
are
( the
) future position of the ballistic missile at tp : x tp and
g cos 𝛾
𝛾̇ = − , y tp for a 2-D scenario. The Input data is composed of 6
V
variables: x0 , y0 , V0 , 𝛾0 , 𝛽  , and tp.
where x, y, V, and 𝛾 indicate downrange, altitude, velocity, Neural network learning is a kind of supervised learn-
and flight path angle of the ballistic target, respectively,
( and
)2 ing, for which a supervisor is required to teach the neural
g is the gravitational constant defined as GME ∕ y + RE network to produce
( correct)outputs. The supervisory data is
where G , ME , and RE are the universal gravitational constant, the data set of 𝐠 tp , t0 , 𝐱0 , 𝛽  , which is prepared by numerical
Earth mass, and Earth mean radius, respectively. The bal- integration of (1) for various values of the input variables.
listic coefficient 𝛽 is related to the ratio of inertia force to Figure 1 shows the training process of the neural network
aerodynamic drag: based on the error feedforward backpropagation algorithm
mg [ ] with the Levenberg–Marquardt optimization. The neural
𝛽= kg∕ms2 . (2) network used in this study has two hidden layers and one
CD S
output layer. The activation function of the hidden layers is
By definition, the ballistic coefficient is infinite if there the hyperbolic tangent sigmoid function. Since the range of
is no drag force. We assume that 𝛽 is constant during the the output variables is not predetermined, a linear function
flight. is chosen as the activation function of the output layer.
Additional to the ballistic coefficient, air density affects The advantage of the neural network application for tar-
aerodynamic forces. Air density varies with temperature get position prediction is that the calculation time is very
Ta , pressure P and altitude as short. It only requires some additions, multiplications and
evaluations of sigmoid functions; if the first hidden layer
𝜌 = P∕RTa , (3) has 15 neuron units and the second has 7 units, the network
where R is specific gas constant (≈ 287.058 J kg−1 K−1). In
the troposphere, where the altitude is below 11 km, the tem-
perature and pressure can be modeled as
Ta = 288.15 − 0.0065y [K],
5.2559 (4)
P = 101325(T∕288.15) [Pa],

and for an altitude above 11 km


Ta = 216 [K],
(5)
P = 22630 exp (−0.00015769(y − 11000)) [Pa].

From (4) and (5), air density can be determined by (3).


Fig. 1  Training of neural network

13
International Journal of Aeronautical and Space Sciences

output is obtained by 209 summations, 209 multiplications, any. In the response surface method, the correction terms,
and 24 activation function evaluations. In applications of the which are the functions of the initial state of the ballistic
trained neural network, the input variables, which are the object, are modeled as linear combinations of the initial
current target states, are estimated from the tracking radar states:
measurements. The ballistic coefficient can also be estimated
𝛿x = ax0 (t0 ) + axx (t0 )x0 + axy (t0 )y0 + axV (t0 )V0 + ax𝛾 (t0 )𝛾0 ,
by the tracking filters such as extended Kalman filter (EKF)
or unscented Kalman filter (UKF). Or a proper value for (11)
the ballistic coefficient can be chosen if the target model is t t
𝛿tF = a0F (t0 ) + atxF (t0 )x0 + atyF (t0 )y0 + aVF (t0 )V0 + at𝛾F (t0 )𝛾0 .
identified. (12)

Remark 1.  Although the neural network shown in Fig. 1 is The time-varying coefficients of (11) and (12) need to be
designed to predict the target position, it is possible to sim- predetermined using the least-square method on the numeri-
ply construct a neural network system to predict the veloc- cal simulation results. First, the reference trajectory of the
ity and the flight path angle by modifying the output of the missile is generated and the time points ( t1 , t2 , … , tN  ), at
neural network. which the coefficients are determined, are selected. The true
range and true flight time at the target altitude for various
initial states given at each time point are obtained by numeri-
2.3 Prediction Using Response Surface Method cal simulations and expressed as

The Kepler problem is to predict the orbit position or veloc- ⎡ ax0 ⎤


⎢ ax ⎥
ity over time given the current states. Based on the solu- � � x
k k k k k k k ⎢ ax ⎥ k k x k
tion of the Kepler’s two-body problem, the algorithm can 𝛿x = x1 − x̃ = 1 x0 y0 V0 𝛾0 ⎢ y ⎥+e =𝐬 𝐚 +e ,
predict the state of a descending ballistic object at a speci- ⎢ ax ⎥
⎢ aVx ⎥
fied altitude (e.g., hit position) [7, 8]. Since the Keplerian ⎣ 𝛾 ⎦
� �
orbit depicts the motion of the body only considering the k = 1, … , Ne ,
gravity, the atmospheric drag effect is not captured in the (13)
algorithm. To compensate the atmospheric drag effect, the
where Ne is the number of numerical simulations, k is the
response surface method developed by Box is used to gener-
simulation number, and e is the modeling residual. Finally,
ate the correction terms for the errors of the impact point and
the coefficients that best represents the correction terms
impact time caused by the drag effect [16, 17]. The response
at each time point can be found by minimizing the sum of
surface method mathematically represents the relationships
squared residuals as follows:
between the input and output variables by employing a sta-
tistical approach. ( )−1
𝐚x = 𝐒T 𝐒 𝐒T Δx , (14)
To briefly introduce the response surface method, we con-
sider a ballistic motion in the two-dimensional Cartesian ( )−1
coordinate system, in which the approximate range and flight 𝐚tF = 𝐒T 𝐒 𝐒T ΔtF , (15)
time are analytically determined as
where
x̃ = −vx0 ̃tF + x0 , (7)
⎛ 𝐬1 ⎞
( √ )/ ⎜ ⎟
( ) 𝐒 = ⎜ ⋮ ⎟, (16)
̃tF = 2
vy0 + vy − 2g yt − y0 g, (8) ⎜ 𝐬Ne ⎟
0
⎝ ⎠
where yt is the target altitude and (x0 , y0 , vx0 , vy0 ) is the cur-
rent state of the object. The true range and true flight time ⎛ 𝛿x1 ⎞ ⎛ 𝛿tF1 ⎞
at the target altitude are related to the approximate values as ⎜ ⎟ ⎜ ⎟
Δx = ⎜ ⋮ ⎟, ΔtF = ⎜ ⋮ ⎟. (17)
⎜ 𝛿xNe ⎟ ⎜ Ne ⎟
x = x̃ + 𝛿x, (9) ⎝ ⎠ ⎝ 𝛿tF ⎠

tF = ̃tF + 𝛿tF , (10) The coefficients at a specific time between two prede-
termined time points can be estimated using the linear
where 𝛿x and 𝛿tF are the correction terms for compensating interpolation method.
the effects of atmospheric drag and other error sources if

13
International Journal of Aeronautical and Space Sciences

3 Ballistic Missile Intercept

Predicted intercept point (PIP) is the point or position where


the interceptor and the target meet each other at the same
time. To determine the PIP, not only the position of the tar-
get at a given time, but the position of the interceptor at the
same time needs to be determined. Prediction of the target
position is implemented using the neural network method
described in Sect. 2. In this section, the algorithm to deter-
mine the PIP is proposed, and midcourse and terminal guid-
ance of the interceptor are described.

3.1 Determination of PIP
Fig. 2  Engagement region for PIP determination
PIP is the final destination of the interceptor during its mid-
course guidance phase, at which the interceptor and the tar-
get arrive at the same time. The interceptor is launched when If the outer boundary of the capture region is preferred for
its flight time to PIP is the same as the remaining flight time PIP, a proper tp satisfying this condition can be found
of the target to PIP. Let tgo(T)
be the flight time of the target through a few iterations. Once a suitable PIP candidate
from its current position to PIP, and tgo (I)
the flight time of the included in the capture region is found, the points of the tgo (I)

interceptor from the launcher to PIP. Then, the interceptor table surrounding the PIP candidate, such as (xi , yi ) shown
needs to wait for tdelay
(I)
determined to satisfy in the figure, {
are found
} and tgo for the PIP candidate is inter-
(I)

polated from tgo, i of these points. If tp > tgo


(I) , it is possible
(T) (I) (I)
tgo = tgo + tdelay . (18) to intercept the target at the PIP candidate by launching the
interceptor after waiting tdelay determined by (18). Otherwise,
Given a candidate for PIP, tgo
(I)
should be calculated quickly
tp should be updated to an alternative one longer, changing
by some means while tgo is easily calculated using the neural
(T)
PIP to a lower altitude.
network trained for target trajectory. For this purpose, three
Note that the PIP prediction procedure is very simple
methods are suggested: (1) tgo (I)
is calculated for a grid of
so that it can be repeated as fast as the cycle of the target
possible PIP’s and tabulated, (2) tgo (I)
for a number of PIP’s
tracking radar. A simple low-pass filter can be employed to
is calculated and the results are learned by another neural
smooth out the noisy sequences of the predicted PIP position
network, and (3) tgo (I)
is calculated by a real-time algorithm.
and tdelay . The preparation of the tgo
(I)
table may require extra
The first and second methods are useful for determining
considerations. Depending on the ambient temperature, the
(I)
tdelay but may not be practical for correcting PIP after the
burn time of solid rocket propellants varies, causing some
interceptor is launched. Contrary to the ballistic target, the variations in tgo
(I)
 . Wind direction and speed could be another
trajectory of the interceptor is not autonomous but affected source of errors in tgo (I)
 . Guidance laws of the launch and
by the guidance activities during its flight. A huge data set midcourse phases assumed for preparation of the tgo (I)
table
is required for neural network training to calculate tgo (I)
in real must be identical to the ones implemented on the intercep-
time for all possible cases of PIP, the current states of the tor missiles.
interceptor, and remaining propellant. The third method is
recommended for tgo (I)
calculation after launch but it is beyond
the scope of this paper. In the following, we assume the first 3.2 Interceptor Configuration and Flight Sequence
method using a pre-calculated table for PIP prediction.
As shown in Fig. 2, the region of intercept is determined It is important to tabulate tgo
(I)
accurately by considering the
in which a number of points densely distributed are selected. dynamics and guidance laws of the interceptor. The flight
Then, tgo (I)
is calculated for each of the points and tabulated. time of the interceptor is determined by the interceptor’s
This preparation procedure is done off-line in advance so the specification such as the launch mass, propellant mass, thrust
calculation time of tgo (I)
dose not matter. level, burn time and guidance law of the booster rocket. In
In the real engagement, an initial guess of tgo (I)
 , denoted as this study, we consider an interceptor vehicle which has a
tp here, is chosen, and the position of the target at T = Tc + tp , single-stage solid-propellant rocket motor and the kill vehi-
where Tc is the current time, is calculated using the neural cle equipped with a liquid-propellant DACS.
network of target trajectory, to check if the target position Let mt be a total mass of the interceptor including m1
belongs to the effective capture region of the defense system. of the rocket motor and m2 of KV; that is, mt = m1 + m2 .

13
International Journal of Aeronautical and Space Sciences

The rocket mass is divided into the propellant mass and the square of the speed V 2 . The aerodynamic coefficients of the
mass of other components such as fuel tank, motor case, interceptor, CL and CD , are assumed as
and thrust vector control devices. Define 𝜈 as the mass ratio
of the structures and other devices to the total rocket mass. CL = CL𝛼 𝛼,
(22)
Then, the mass of propellant is written as mp = (1 − 𝜈)m1 . CD = CD0 + kD CL2 ,
If the specific impulse of the solid propellant Isp and the
maximum allowable axial acceleration nmax are given, the where CL𝛼 and CD0 are constant coefficients, 𝛼 = 𝜃 − 𝛾I is
rocket thrust F and the burn-time tburn are determined as: the angle of attack of the interceptor, and kD is the induced
drag coefficient.
F = mbo g nmax , Adopting the approach of [20, 21], we optimize the boost
mp g Isp (19) phase trajectory of the interceptor employing a practical guid-
tburn = ,
F ance law. This approach has a great advantage over a rigorous
trajectory optimization since the optimized flight time can be
where mbo is the burn-out mass of the interceptor expressed
realized accurately in applications. The guidance law used for
as mbo = 𝜈m1 + m2.
the boost phase here is the linear tangent law (LTL), which is
The entire flight of the interceptor proceeds in the fol-
known as the optimal solution to the ideal ascending problem
lowing order: At first, the interceptor is launched vertically
of a launch vehicle [18]. The pitch angle command of LTL for
and propelled to the designated PIP by the rocket motor.
the boost phase is given as
The rocket is separated right after its burn-out time, and KV
( )
enters an un-powered midcourse guidance phase in which a t − tign + 𝜋∕2
the trajectory is frequently corrected by DACS. This correc- tan 𝜃 ≈ 𝜃 = ( ) , (23)
b t − tign + 1
tion is based on the target information provided by a ground
tracking radar and the PIP update procedure proposed in where, tign is the ignition time of the rocket motor, and the
this paper. When the on-board IR seeker detects and tracks constants a and b are guidance parameters to be optimized.
the target, KV enters its terminal homing guidance phase to Other
achieve a hit-to-kill. ( )constants is set to satisfy the vertical launch condition,
𝜃 tign = 90◦.
The objective of the optimization problem is to find guid-
3.3 Boost Phase Guidance Law of Interceptor ance parameters a and b that minimize the flight time tf to
deliver the interceptor to the designated PIP. The optimization
In the boosting phase, the equations of motion of the inter- problem is then written as
ceptor in the two-dimensional Cartesian coordinate system
is described below. min J = tf ,
a,b (24)
ẋ I = Vx ,
subject to (20), (23), and the initial and final conditions of
ẏ I = Vz , the interceptor. While the final speed
( is free,
) the final posi-
F cos 𝜃 − L sin 𝛾I − D cos 𝛾I tion is one of the PIP candidates, xi , yi  , shown in Fig. 2.
V̇ x = ,
m In this optimization procedure, KV is assumed to follow a
(20)
F sin 𝜃 + L cos 𝛾I − D sin 𝛾I ballistic trajectory from the booster burn-out position to the
V̇ y = − g,
m selected PIP without any maneuver. The above optimiza-
F tion problem is solved using the co-evolutionary augmented
ṁ = − ,
Isp g Lagrangian method (CEALM) [19], which is an evolution-
ary method for solving constrained parameter optimization
where L , D , and 𝜃 are the lift force, drag force and pitch problems (Fig. 3). The optimized values of tgo (I)
for the set
angle, respectively, and the(flight path
) angle of the intercep- of PIP candidates and the interceptor model described in
tor is defined as 𝛾I = tan−1 Vy ∕Vx  . The aerodynamic forces Sect. 4.2 are illustrated in Fig. 4. In addition to the flight
are given as: time for each PIP candidate, the LTL parameters, a and b ,
L = CL Q S, are also tabulated for the boost phase guidance. The table of
(21)
(I)
tgo is used to determine the launch time while the table of a
D = CD Q S,
and b are interpolated for boost-phase guidance.
where S is the reference area of the body of the missile and
Q is the dynamic pressure expressed as Q = (1∕2) 𝜌 V 2 . The
dynamic pressure is proportional to the air density 𝜌 and the

13
International Journal of Aeronautical and Space Sciences

homing guidance law may not be able to reduce the accumu-


lated heading errors to zero, leading to interception failure.
Correction of the KV trajectory requires four functions:
(1) prediction of the remaining part of target trajectory, (2)
prediction of the remaining part of interceptor trajectory, (3)
PIP update, and (4) trajectory correction using DACS. The
first function can be done by applying the neural network
method to the latest target information provided the ground
radar. The second function can also be processed using a
neural network dedicated to the interceptor trajectory predic-
tion but an analytical solution based on the quadratic drag
model is employed. PPI update is required if the target and
the interceptor do not arrive at the current PIP at the same
time. A simple iterative algorithm has been developed for
this purpose. The last function of trajectory correction is
done by controlling the vertical component of the velocity
vector to produce the required change in altitude at the pre-
dicted intercept time.
Let Tc and tgo,k be the current time and the current time
to go. Then, we denote the interceptor position and velocity
at T = Tc + tgo,k as 𝐱Mk and 𝐯Mk , respectively. Similarly, the
target position and velocity at T = Tc + tgo,k are denoted as
Fig. 3  Flow chart of PIP determination 𝐱Tk and 𝐯Tk , respectively. As explained earlier, 𝐱Mk and 𝐯Mk
are calculated using the analytical solution studied in [22,
23] while 𝐱Tk and 𝐯Tk are predicted using the neural network
trained for target trajectory. Suppose that the current time
to go needs to be adjusted by Δt to minimize the miss dis-
tance. Then, the positions of the interceptor and the target
are given as
𝐱Mk+1 = 𝐱Mk + 𝐯Mk Δt,
(25)
𝐱Tk+1 = 𝐱Tk + 𝐯Tk Δt,

and the position error at T = Tc + tgo,k + Δt is obtained as


E = 𝐱Mk+1 − 𝐱Tk+1 . Now we solve for Δt∗ , which minimizes
JE = ET E , by applying the necessary condition 𝜕JE ∕𝜕Δt = 0.
A simple calculation produces the optimal solution as
( )T ( )T
𝐯Mk − 𝐯Tk 𝐱Mk − 𝐱Tk
Δt∗ = − ( )T ( ). (26)
𝐯Mk − 𝐯Tk 𝐯Mk − 𝐯Tk
Fig. 4  Interceptor flight times to PIP candidates
Then, the time-to-go is updated as tgo,k+1 = tgo,k + Δt∗ and
3.4 Trajectory Correction in Midcourse Phase PIP is updated by 𝐱Tk+1;

During the midcourse phase, the KV trajectory is controlled PIPk+1 = 𝐱Tk + 𝐯Tk Δt∗ . (27)
by a liquid-propellant DACS. If the PIP solution is the exact
A few iterations are required if the position error at the
intercept point, KV does not need to correct the trajectory. In
updated PIP is to be reduced to a negligible magnitude.
reality, however, the PIP solution is prone to some errors due
However, the refinement of PIP is not necessary since the
to various reasons; target information may not be accurate,
PIP update will be conducted repeatedly during the mid-
the missile dynamic model used for tgo (I)
calculation may not
course phase.
be accurate, or the actual launch time may be different from
If the amount of PIP correction is not negligible, the
the designated launch time. Without correction of the KV
interceptor trajectory should be corrected so that the
trajectory of KV during the midcourse phase, the terminal

13
International Journal of Aeronautical and Space Sciences

interceptor arrives at the updated PIP at the updated inter- Table 2  Neural network test cases
cept time. A simple way for this correction is to produce x (km) y(km) V(km/s) 𝛾(deg) 𝛽(kg/ms2 )
a velocity component of the interceptor to the direction
perpendicular to the velocity vector. The amount of the Case 1 120 60 3.5 − 30 20,000
required lateral velocity component is calculated as Case 2 130 60 3.5 − 30 50,000
Case 3 120 100 2.0 − 40 30,000
ΔV⊥ = ΔPIP⊥ ∕tgo,k+1 . (28) Case 4 130 100 2.0 − 40 60,000

where ΔPIP⊥ is the component of the PIP change perpen-


dicular to the current interceptor velocity vector. Since the
flight path angle of the interceptor does not vary significantly 4 Numerical Experiments
during the midcourse phase, the velocity correction of (34)
is quite effective as found in the numerical experiments. 4.1 Accuracy of Target Trajectory Neural Network

In the numerical study, the launch site of the interceptor is


3.5 Terminal Guidance Law of Interceptor chosen as the origin of the coordinate system and the initial
position of the ballistic target is a point of positive range
The guidance law employed for the terminal homing phase and altitude. The input variables of the neural network to be
is an augmented proportional navigation (APN) guidance trained to learn target trajectory are composed of 6 variables:
law, which is true proportional navigation (TPN) aug- current range, current altitude, current speed, current flight
mented by two additional terms to compensate for the tar- path angle, ballistic coefficient, and prediction time interval.
get acceleration and gravity. The guidance law is written as The values of the input variables are bounded as shown in
Table 1 for the generation of the supervisory data.
N( ) A total of 230,000 supervisory trajectory data are gen-
aM = NVc 𝜆̇ + g cos 𝛾I + aT , (29)
2 erated by random selection of the input variables within
where N is the navigation ratio between 3 and 5, Vc is the their bounds and numerical integration of the equations
closing velocity, 𝜆̇ is the line-of-sight rate, 𝛾I is the flight of motion. All trajectory portions of negative altitude are
path angle of the interceptor, and aT is the target acceleration deleted from the data set. A neural network of 15 and 7
component perpendicular to the line of sight. The direction neurons for the two hidden layers is then trained to output
of divert thrust is assumed to be perpendicular to the line of the terminal target position of the trajectory data for about
sight. In this study, we also assume that the target accelera- 20,000 epochs.
tion is precisely estimated by an on-board target tracking As a reference to evaluate the prediction accuracy of
filter. If the interceptor is equipped with an efficient target the trained neural network, the target trajectory is gener-
tracking filter, various optimal guidance laws such as pro- ated using the Runge–Kutta 4th-order method for 4 initial
posed in [24, 25] are also applicable to homing guidance
of KV.

100

case 1

80 case 2
case 3
case 4

60
Altitude (km)

Table 1  Bounds of neural network input variables 40

Variables Unit Lower Upper


20
x km 50 150
y km 50 120
V km/s 1 5 0
20 40 60 80 100 120 140
𝛾 deg − 70 − 10
Downrange (km)
𝛽 kg/m ⋅ s2 5000 80,000
tp sec 0.1 65
Fig. 5  Reference trajectory

13
International Journal of Aeronautical and Space Sciences

250 Table 3  Number of neurons for hidden layers


case1: = 20000
case2: = 50000
Name of Number of neurons in
200 case3: = 30000 Neural network Hidden layer 1 Hidden layer 2
case4: = 60000
N1 20 10
150
N2 15 7
Position Error (m)

N3 10 7
100

350

50
300 N1
N2
N3
0 250
0 10 20 30 40 50 60 70
Predicted Time (sec)
200

Position Error (m)


Fig. 6  Neural network prediction errors 150

100

50

0
0 10 20 30 40 50
Predicted Time (sec)

Fig. 8  Prediction errors of three neural networks

(larger drag). At low altitudes, the target speed suddenly


decreases at a very high rate as the air density increases. For
this reason, the relationship between the prediction time tp
and the target states becomes highly nonlinear. That explains
why the prediction errors are relatively large when the target
altitude is low at tp.
Fig. 7  Prediction errors for various ballistic coefficients The number of neurons used in the hidden layers is impor-
tant for the trajectory prediction errors. As shown in Table 3,
three neural networks using different numbers of neurons are
conditions given in Table 2. The reference trajectories are trained to learn the target trajectory data. Figure 8 displays
shown in Fig. 5, where we observe significant influence of that the neural network with the most neurons provides the
the ballistic coefficient on the trajectory shape at low alti- smallest prediction errors. The average position error of each
tudes. (A smaller ballistic coefficient implies larger aerody- neural network is 8.82 m, 56.82 m, and 96.48 m, respec-
namic drag.) tively. This experiment implies the target trajectory can be
Figure 6 shows the prediction errors of the neural network learned well even for low altitudes if the number of neurons
predictor along with the four cases of the reference trajec- for each hidden layer is judiciously chosen.
tory. The average prediction errors are at most 60 m for all
cases. It is noted that the prediction errors increases at low 4.2 ABM Engagement Simulation
altitudes where the effect of aerodynamic drag is significant.
To check the influence of the ballistic coefficient on pre- Two-dimensional anti-ballistic missile (ABM) engage-
diction errors, four reference trajectory samples are chosen ments are simulated using the neural network N2 described
again. The initial conditions except 𝛽 are the same as those in Table 3 with the data table of interceptor’s flight times.
of Case 4. As observed in Fig. 7, the prediction errors at high The specification of the interceptor are given in Table 4.
altitudes (short prediction times) are not much dependent The burn time of the solid propellant and thrust are deter-
on the magnitude of the ballistic coefficient but larger pre- mined by the specification. The navigation constant of
diction errors are produced for smaller ballistic coefficients terminal guidance is chosen as 5. It is assumed that the

13
International Journal of Aeronautical and Space Sciences

Table 4  Specification of the Booster KV demonstrate the usefulness of the neural network method
interceptor for PIP determination.
Diameter [m] 0.25 0.25
Mass [kg] 520 75 4.3 Effects of PIP Correction During the Midcourse
𝜈 0.18 – Phase
Thrust [N] 56,235 –
Burn time [s] 18.6 – In real engagements, there are various uncertainties such as
Isp [s] 260 – errors in the aerodynamic models and thrust model. Under
CD0 0.2 0.2 these uncertainties, the interceptor’s flight time to the des-
CL𝛼 4.5 1.5 ignated PIP is affected. For example, if the actual thrust is
kD 0.5 0.5 smaller than the model, the interceptor would be delivered
to the designated PIP later than expected. To investigate the
effects of uncertainties, the launch time TL is intentionally per-
ground-based tracking radar provides the target states turbed to be a little earlier or later. Note that a small launch
every 1 s and PIP calculation is repeated at the same fre- time error could produce a huge heading error at the beginning
quency. PIP update with a higher rate is possible since it of the homing phase due to the very high relative velocity
takes only 0.05 s on average in this numerical experiment. between the target and the kill vehicle.
The initial conditions of the target are given in Table 5. With the initial conditions of Scenario 1 given in Tables 5
Determination of tgo(T)
starts from its initial guess shown in and 6, the interceptor launch time is changed from T = 16 s to
the last column of Table 5. If tp is not proper to be tgo
(T)
 , then T = 15 s and T = 17 s, respectively, as shown in Table 8.
tp is updated and the same procedure is repeated at the next In Scenario 1–1, the intercept point is changed to a lower
cycle. Suppose that T = 0 at the initial time and T = TPIP at altitude due to the launch delay of 1 s, as shown in Fig. 10a
the time of PIP determination. By definitions, we see that and downward guidance commands are generated as shown
(I)
in Fig. 10b. In Scenario 1–2, we have the opposite results as
Tlaunch = TPIP + tdelay , (30) shown in Fig. 11. It is observed that if PIP is not corrected
during the midcourse phase, large initial heading errors of the
(I) (T) homing phase results in excessive guidance commands and
Tintercept = Tlaunch + tgo = TPIP + tgo . (31)
drastic reduction of KV’s velocity. If KV is not able to produce
Figure 9 shows the interceptor/target trajectory, zero such large lateral accelerations, the guidance performance
effort miss, guidance commands of the interceptor, and could be significantly deteriorated.
interceptor velocity for Scenario 1. The terminal homing
phase takes only 7 s due to a high closing velocity. The 4.4 Comparison with the Response‑Surface Method
zero effort miss approaches to 0 at the final time as shown (RSM)
in Fig. 9b.
Table 6 compares the time of PIP determination, PIP In this section, the proposed algorithm, which is based on the
position, time to go of the target to PIP, launch time of neural network method, is compared with an IIP calculation
the interceptor, and time of intercept of the two scenarios. algorithm using RSM treated in [7]. RSM calculates the target
In Scenario 1, the initial guess of tgo (T)
is 55 s, which is state errors with respect to the analytic Keplerian solution as
a proper guess. In Scenario 2, the initial guess of tgo (T)
is linear functions of the initial conditions. To apply RSM to PIP
40 s, which is not enough for the target to reach the cap- determination, we need to specify the altitude of the intercep-
ture region. Nonetheless, we observe that the intercept tion in advance. If multiple values of altitudes are necessary,
time, TI  , is not much sensitive to the choice of tp . Table 7 the coefficients of Eq. (13) should be prepared for each inter-
compares the predicted values with the simulation results cept altitude. To apply RSM algorithm, the intercept altitude
for the impact point and intercept time. Small differences is required to predict the target position. Details of the RSM
between the predicted values and the true ones clearly algorithm can be found in [7].
Figure 12 and Table 9 show the performance of the RSM
and NN methods for Scenario 1. Both methods calculate PIP

Table 5  Initial conditions of x (km) y (km) V (km/s) 𝛾 (deg) 𝛽 (kg/ms2 ) tp (sec)


scenarios
Scenario 1 120 100 1.5 − 40 20,000 55
Scenario 2 120 100 1.5 − 40 20,000 40

13
International Journal of Aeronautical and Space Sciences

(a) (b)

(c) (d)

Fig. 9  Scenario 1: a trajectory, b zero effort miss, c terminal guidance command, d velocity

Table 8  Launch time of each Scenario # TL (s)


Table 6  Engagement simulation results scenario
Scenario 1 (nominal) 16
Scenario # TPIP (sec) PIP(km) (T)
tgo (s) TL (s) TI (s)
Scenario 1–1 17
Scenario 1 0 (56.86, 32.57) 55.0 16.0 55 Scenario 1–2 15
Scenario 2 5 (57.20, 32.98) 49.7 15.5 54.7

successfully and produce satisfactory guidance performance.


While the computation of the coefficients of RSM is less time
Table 7  Comparison of predicted and true values
consuming than NN, RSM has a drawback that the coefficients
for various combinations of initial target altitude and intercept
Scenario # Intercept point (km) Intercept time (s) altitude should be prepared for practical application. Other-
Predicted True Predicted True wise, the calculation of PIP will not be possible whenever
the altitude of PIP does not match with any of the intercept
Scenario 1 (56.86, 32.57) (56.99, 32.72) 55.00 54.94
altitudes considered in RSM. On the other hand, PIP calcula-
Scenario 2 (57.20, 32.98) (57.32, 33.15) 54.70 54.64
tion using neural network is free from this drawback since

13
International Journal of Aeronautical and Space Sciences

(a) (b) (c)

Fig. 10  Comparison of Scenarios 1 and 1–1: a trajectory, b guidance command, c velocity

(a) (b) (c)

Fig. 11  Comparison of Scenarios 1 and 1–2: a trajectory b guidance command, c velocity

(a) (b) (c)

Fig. 12  Comparison of NN and RSM: a trajectory b guidance command, c velocity

13
International Journal of Aeronautical and Space Sciences

Table 9  Comparison of NN and RSM AIAA Guidance, Navigation, and Control Conference and Exhibit,
Austin, TX. AIAA Paper 2003–5722.
Intercept position (km) Intercept time (s) 7. Ahn J, Roh WR (2012) Noniterative instantaneous impact point
Predicted True Predicted True prediction algorithm for launch operations. J Guidan Control Dyn
25:2
NN (56.86, 32.57) (56.99, 32.72) 55.00 54.94 8. Zarchan P (2002) Tactical and strategic missile guidance, Vol.
199, Progress in astronautics and aeronautics, AIAA, Reston
RSM (59.53, 36.00) (59.58, 36.06) 52.69 53.02
9. Jeon IS, Lee JI, Tahk MJ (2010) Homing guidance law for coop-
erative attack of multiple missiles. J Guid Control Dyn 33:136
10. Kim TH, Lee CH, Tahk MJ (2011) Time-to-go polynomial guid-
the neural network learns continuous functions associating the ance laws with terminal impact angle acceleration constraints.
current states with the future states. Proceedings of the ­18th IFAC World Congress, Milano
11. Lee CH, Kim TH, Tahk MJ (2013) Interception angle control
guidance using proportional navigation with error feedback. J
Guid Control Dyn 36:5
5 Conclusion 12. Tahk MJ, Ryoo CK, Cho H (2002) Recursive time-to-go estima-
tion for homing guidance missiles. IEEE Trans Aerosp Electron
Syst 38:1
This work proposes a predicted intercept point (PIP) deter- 13. Whang IH, Ra WS (2008) Time-to-go estimator for missiles
mination algorithm based on neural network learning for guided by BPNG. In: Proceedings of IEEE International Confer-
anti-ballistic missile defense systems. A neural network ence on control, automation and systems, IEEE Publ., Piscataway,
is trained to learn the functional relationship of the future NJ.
14. Ghosh S, Ghose D, Raha S (2013) Three dimensional retro-PN
target position with the current time and the prediction based impact time control for higher speed nonmaneuvering tar-
time interval. The trained neural network enables precise gets. In: Proceedings of the IEEE Conference on Decision and
PIP calculation to be done within the order of 0.1 s. Such Control, IEEE Publ., Piscataway, NJ
computation speed will be very helpful for realistic anti- 15. Dhananjay N, Ghose D (2014) Accurate time-to-go estimation for
proportional navigation guidance. J Guid Control Dyn 37:4
ballistic missile engagements. Although it is not pursued in 16. Box G, Wilson K (1951) On the experimental attainment of opti-
this work, another neural network trained for the flight time mum conditions. J R Stat Soc Ser B 13:1
of the interceptor from the current position to various inter- 17. Ahn J, Seo J (2012) Instantaneous impact point prediction using
cept points could shorten the PIP calculation time further. the response surface method. J Guid Control Dyn 35(2):645–648
18. Bryson AE (1975) Applied optimal control: optimization, estima-
Since the interceptor trajectory may have three-dimensional tion and control. CRC Press, Boca Raton
shapes, development of a neural network method requires 19. Tahk MJ, Sub BC (2000) Coevolutionary augmented Lagrangian
innovative ideas to reduce the size of trajectory data to be methods for constrained optimization. IEEE Trans Evol Comput
learned. 4(2):114–124
20. Lee JY, Hong SM, Kim YW, Tahk MJ (2015) Parameter optimi-
zation of multistage missile for terminal velocity maximization.
Acknowledgements  This work was conducted at High-Speed Vehicle Asia-Pacific International Symposium on Aerospace Technology,
Research Center of KAIST with the support of Defense Acquisition Engineers Austrailia
Program Administration (DAPA) and Agency for Defense Develop- 21. Hong SM, Tahk MJ, Lee CH (2019) Stage optimization of anti-
ment (ADD) (contract number: UD170018CD). air missiles using practical guidance laws. Int J Aeronaut Sp Sci,
1–10.
22. Chudinov PS (2011) Approximate analytical investigation of pro-
References jectile motion in a medium with quadratic drag force. Int J Sports
Sci Eng 5(1):27–42
23. Han S, Hwang MC, Lee BY, Ahn J, Tahk MJ (2016) Analytic
1. Zarchan P (1999) Ballistic missile defense guidance and control solution of projectil motion with quadratic drag and unity thrust.
issues. Sci Glob Secur 8:1 IFAC-PapersOnLine 49(17):40–45
2. Barton DK, Falcone R, Kleppner D, Lamb FK, Lau MK, Lynch 24. Kim TH, Lee CH, Jeon IS, Tahk MJ (2013) Augmented poly-
HL, Moncton D, Montague D, Mosher DE, Priedhorsky W, Tigner nomial gduiacne with impact time and angle constraints. IEEE
M, Vaughan DR (2004) Report of the American physical society Transactions on aerospace and electronics systems, Vol. 49, No.
study group on boost-phase intercept systems for national missile 4.
defense: scientific and technical issues. Rev Modern Phys 76:1307 25. Seo MG, Tahk MJ (2017) Suboptimal mid-course guidance algo-
3. Zhang X, Lei H, Li J, Zhang DY (2014) Ballistic missile trajectory rithm for accelerating missiles. In: Proceedings of the Institution
prediction and the solution algorithms for impact point prediction. of Mechanical Engineers, Part G: Journal of Aerospace Engineer-
In: Proceedings of 2014 IEEE Chinese Guidance, Navigation and ing 31(11)
Control Conference, Yantai, China
4. Song EJ, Tahk MJ (1998) Real-time midcourse guidance with
Publisher’s Note Springer Nature remains neutral with regard to
intercept point prediction. Control Eng Pract 6(8):41
jurisdictional claims in published maps and institutional affiliations.
5. Montenbruck O, Markgraf M, Jung W, Bull B, Engler W (2002)
GPS based prediction of the instantaneous impact point for sound-
ing rockets. Aerosp Sci Technol 6:962
6. Markgraf M, Montenbruck O, Turner P, Viertotak M (2003)
Tracking system with onboard IIP prediction for sounding rockets,

13

You might also like